Quantization, distillation, pruning and serving work aimed at the same accuracy for less memory, latency and money.
Embodied Reasoning constitutes a fundamental capability of embodied intelligence, serving as the basis for autonomous perception, reasoning, and interaction within physical environments. Recent studie...
Multimodal Large Language Models (MLLMs) are increasingly deployed as multi-step agents, where explicit reasoning supports task decomposition and tool coordination but also accumulates self-generated...
Large-scale rollouts have become a core component of modern LLM systems, spanning reinforcement learning (RL) post-training, on-policy distillation (OPD), and sampling-heavy evaluation pipelines. Unli...
arXiv:2606.30159v2 Announce Type: replace Abstract: Dual-energy CT (DECT) exploits attenuation differences across different X-ray spectra to provide richer material information and has been widely us...
arXiv:2608.21019v1 Announce Type: cross Abstract: Quantization is widely used to deploy large language models, but its effect on uncertainty behavior, such as confidence, margins, and abstention, is...
ForeTime‑VLA is a causal vision‑language‑action policy that distills future‑aware representations from a frozen Fast‑WAM teacher, enabling it to anticipate contact events during conveyor‑belt manipulation. The method compresses current and future video latents into a 64‑dimensional target, uses an eight‑frame history encoder to predict this target along with manipulation phase and time‑to‑transition, and conditions a VLM prefix on future tokens and phase. On a deduplicated conveyor‑belt dataset, ForeTime‑VLA reduces test MAE by 2.63% and L2 by 3.02%, while real‑robot experiments show significantly higher grasp success rates compared to the next‑best reference. whyItMatters":"The approach demonstrates that distilling future‑token knowledge from a world‑action model can improve dynamic manipulation performance without the computational cost of running the teacher at inference time."
arXiv:2608.20554v1 Announce Type: cross Abstract: The critical failure modes in deployed large language models (LLMs) are cross-dimensional: a model can score 99.3 in safety alignment while refusing...
arXiv:2608.21343v1 Announce Type: cross Abstract: Contextualization is essential for production automatic speech recognition (ASR) systems, where user-provided phrases must be recognized accurately u...
MentorPulse introduces a method for refreshing cross‑model latent guidance during long‑form generation. By compressing mentor states into a capped slot memory and updating it incrementally as new tokens are generated, the approach keeps the guidance signal fresh without resetting the student’s KV cache. Across thirteen datasets, MentorPulse closes more than half of the mentor‑student performance gap, outperforming existing techniques such as C2C, T2T, and LoRA, especially on long outputs.
arXiv:2606.18250v2 Announce Type: replace Abstract: Forecasting the evolution of dynamic environments is crucial for autonomous agents. While generative world models have achieved high photorealism i...
arXiv:2511.17886v2 Announce Type: replace-cross Abstract: Vision-language models (VLMs) have achieved remarkable success across multimodal tasks, yet their substantial computational demands hinder ef...
arXiv:2608.20375v1 Announce Type: new Abstract: Tree-based speculative decoding raises the mean accepted tokens of standard speculative decoding by verifying multiple draft paths, and existing tree b...
arXiv:2608.20953v1 Announce Type: cross Abstract: Serving large language models cheaply increasingly means shipping models that are both structurally compressed to a fraction of their parameters and...
TreeWY introduces a speculative verification method for Gated DeltaNet (GDN) hybrid models that eliminates the need for per-draft-state snapshots. By applying a tree‑structured WY transform to the gated delta rule, each draft node’s output is computed with a single triangular solve, and only the accepted state is reconstructed on commit. Benchmarks on Qwen3.5 35B and 397B show reduced memory pressure, higher throughput, and lower time‑to‑first‑token in memory‑bound scenarios, while enabling wider, higher‑acceptance draft trees.
TRACE is a training-time framework that uses structured radiology reports to guide concept editing, allowing image-only diagnosis during inference. It refines image-derived concepts with a teacher-guided editing mechanism in a malignancy-aware ordered concept space and introduces Strategic Concept Missing Training to handle incomplete annotations. The authors also present BUSC, a benchmark linking images, labels, and structured attributes, and show that TRACE outperforms existing methods on multiple datasets with better cross-domain robustness.
arXiv:2608.21030v1 Announce Type: cross Abstract: Video multimodal large language models have advanced significantly, yet fine-grained motion-temporal understanding remains fragile. The core bottlene...
The paper investigates whether distilling reasoning traces from large teacher models is worth the extra compute compared to standard instruction fine‑tuning (IFT). By generating paired IFT and reasoning outputs from the same teacher and training student models at five scales, the authors find that, at matched FLOPs, IFT generally lies on or near the Pareto frontier across most configurations. Reasoning traces only reach the frontier on open‑ended tasks for models 7B and larger, and a curriculum mixing 25–50% reasoning data with IFT can capture most of the accuracy benefit at a lower compute cost.
arXiv:2608.21308v1 Announce Type: new Abstract: Test-Time Training (TTT) enables long-context processing via continuous weight updates during inference, but current methods struggle to balance the ex...
The paper identifies a specific issue in supervised fine‑tuning (SFT) of large language models called factual access failure, where models can recognize correct facts under constrained tests but fail to generate them in open‑ended settings. It demonstrates that SFT can cause both genuine wrong answers and expression‑level errors such as verbosity or formatting mismatches. To mitigate this, the authors propose Recall‑Anchored Distillation (RAD), a self‑distillation method that aligns the fine‑tuned model with the base model’s soft output distribution on unlabeled out‑of‑distribution text, thereby recovering lost factual recall without needing labeled data.
arXiv:2608.20988v1 Announce Type: cross Abstract: Quantization of Large Language Models (LLMs) is often hindered by the sensitivity of the self-attention mechanism to discretization errors. We identi...